{"id":48448498,"url":"https://github.com/thoeltig/file-format-token-accuracy-benchmark-results","last_synced_at":"2026-04-06T19:02:10.325Z","repository":{"id":341489583,"uuid":"1169592172","full_name":"thoeltig/file-format-token-accuracy-benchmark-results","owner":"thoeltig","description":"Benchmark results measuring token efficiency and accuracy across file formats for LLM consumption. Archive of data, reports, and metrics comparing CSV, JSON,   YAML, TOON and more.","archived":false,"fork":false,"pushed_at":"2026-04-01T00:26:26.000Z","size":2884,"stargazers_count":0,"open_issues_count":2,"forks_count":0,"subscribers_count":0,"default_branch":"develop","last_synced_at":"2026-04-01T02:45:00.078Z","etag":null,"topics":["benchmark","benchmark-results","claude","data-formats","file-formats","llm","research","token-efficiency"],"latest_commit_sha":null,"homepage":"","language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/thoeltig.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-02-28T23:02:46.000Z","updated_at":"2026-03-06T02:39:56.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/thoeltig/file-format-token-accuracy-benchmark-results","commit_stats":null,"previous_names":["thoeltig/file-format-token-accuracy-benchmark-results"],"tags_count":1,"template":false,"template_full_name":null,"purl":"pkg:github/thoeltig/file-format-token-accuracy-benchmark-results","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/thoeltig%2Ffile-format-token-accuracy-benchmark-results","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/thoeltig%2Ffile-format-token-accuracy-benchmark-results/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/thoeltig%2Ffile-format-token-accuracy-benchmark-results/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/thoeltig%2Ffile-format-token-accuracy-benchmark-results/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/thoeltig","download_url":"https://codeload.github.com/thoeltig/file-format-token-accuracy-benchmark-results/tar.gz/refs/heads/develop","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/thoeltig%2Ffile-format-token-accuracy-benchmark-results/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":31485516,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-04-06T17:22:55.647Z","status":"ssl_error","status_checked_at":"2026-04-06T17:22:54.741Z","response_time":112,"last_error":"SSL_read: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["benchmark","benchmark-results","claude","data-formats","file-formats","llm","research","token-efficiency"],"created_at":"2026-04-06T19:02:09.126Z","updated_at":"2026-04-06T19:02:10.315Z","avatar_url":"https://github.com/thoeltig.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"# File Format Token Accuracy Benchmark Results\n\nArchive of benchmark results measuring token usage and retrieval accuracy across file formats for LLM consumption to find the most efficient format.\n\n## Overview\n\nThis repository contains **benchmark results and raw data** from experiments evaluating which file formats deliver the most reliable information to LLMs with optimal token efficiency.\n\n**To run your own benchmarks**, see the [benchmark plugin repository](https://github.com/thoeltig/file-format-token-accuracy-benchmark).\n\n## Initial Benchmark Results\n\n- **Date**: December 19, 2025\n- **Model**: Claude 4.5 Haiku\n- **Extended Thinking**: Off\n- **Tested Formats**: 7 (CSV, JSON Compact/Pretty, JSONL, TOON, Markdown, YAML)\n- **Data Variants**: Mandatory and optional fields, 40 \u0026 80 record datasets\n\n### Key Findings\n\n- **CSV**: Unbeatable for dense, mandatory data (70.98% weighted accuracy @ 9,008 tokens). Accuracy collapses ~15% with sparse data.\n- **JSON Compact**: Recommended baseline (70.12% weighted accuracy, 15,957 tokens). Only format that uses fewer tokens with optional fields.\n- **YAML**: Highest accuracy (71.96% weighted) but at 2.62x token cost compared to CSV.\n- **Markdown**: Catastrophic failure (24.67% weighted accuracy despite token efficiency). Unreliable format.\n- **TOON**: Competitive on dense data but token cost explodes 2.17x with optional fields.\n\n### Format Recommendations (Summary)\n\n| Scenario | Format | Notes |\n|----------|--------|-------|\n| **Default choice** | JSON Compact | Best balance of accuracy and token efficiency |\n| **Dense, complete data** | CSV | Unbeatable tokens/accuracy ratio (9,008 tokens @ 71% weighted) |\n| **Maximum accuracy** | YAML | 71.96% weighted, but 2.62x token cost vs CSV |\n| **Sparse/optional fields** | JSON Compact | Only format reducing tokens with optional data |\n| **Never use** | Markdown | 24.67% accuracy = 76% token waste |\n\n**See [Full Report](./initital_benchmark_haiku_4_5_formats_all_variants_all_extended_thinking_off/BENCHMARK_REPORT.md) for detailed decision framework and trade-off analysis.**\n\n## Repository Structure\n\n### Benchmark Runs\n\nEach benchmark run directory contains:\n\n```\nbenchmark/\n├── BENCHMARK_REPORT.md              # Comprehensive analysis and findings\n├── data/                            # Raw test data files in all formats\n│   └── {format}/\n│       ├── *.csv, *.json, *.yaml    # Data files tested\n├── questions/                       # Test question sets\n├── answers_validation/              # Expected answers for validation\n├── subagent_outputs/                # Raw agent responses (3 runs per variant)\n├── results/                         # Validation results per format\n├── metadata.json                    # Dataset characteristics\n├── metrics.json                     # Token and accuracy metrics\n└── analytics_results.json           # Final rankings and insights\n```\n\n**Key Files per Benchmark**:\n- `BENCHMARK_REPORT.md` - Complete analysis, recommendations, and methodology\n- `metrics.json` - Token efficiency and accuracy metrics (all formats)\n- `analytics_results.json` - Efficiency rankings and comparative insights\n- `data/` - Raw data files used in testing\n- `subagent_outputs/` - Raw LLM responses for reproducibility\n\n## Running Benchmarks\n\nTo generate new benchmark results:\n\n1. Clone the [benchmark plugin repository](https://github.com/thoeltig/file-format-token-accuracy-benchmark)\n2. Follow setup instructions there\n3. Run `/benchmark` command with desired parameters\n4. Results are generated in the specified output directory\n\n## Data File Structure\n\nEach benchmark generates standardized test data:\n\n- **60 records** (standardized across benchmarks)\n- **22 fields** (19 mandatory + 3 optional)\n- **Product dataset** (consistent across formats for fair comparison)\n- **Multiple variants**:\n  - Flat and nested structures\n  - Mandatory fields only\n  - Sparse data (optional fields)\n\nAll data is formatted identically in each file format to ensure apples-to-apples comparison.\n\n## Methodology\n\n### Evaluation Criteria\n\n1. **Token Efficiency** - Total tokens consumed per format\n2. **Information Accuracy** - Weighted accuracy (66.7% retrieval/structure, 33.3% filtering/aggregation)\n3. **Consistency** - Performance across data variants (mandatory vs optional fields)\n4. **Robustness** - Scaling behavior with data volume\n\n### Question Weighting\n\n- **Field Retrieval (37.5%)** - Extract specific values\n- **Structure Awareness (29.2%)** - Understand data organization\n- **Filtering (20.8%)** - Count matching criteria\n- **Aggregation (12.5%)** - Sum/average across records\n\nWeighted accuracy prioritizes **understanding data organization** over model reasoning capability.\n\n---\n\n## License\n\nSee root [LICENSE](./LICENSE) for details.\n\n## Related\n\n- **Benchmark Plugin**: [file-format-token-accuracy-benchmark](https://github.com/thoeltig/file-format-token-accuracy-benchmark)\n- **Author**: [Thore Höltig](https://github.com/thoeltig)","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fthoeltig%2Ffile-format-token-accuracy-benchmark-results","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fthoeltig%2Ffile-format-token-accuracy-benchmark-results","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fthoeltig%2Ffile-format-token-accuracy-benchmark-results/lists"}